Exploring LLM Priming Strategies for Few-Shot Stance Classification

How does priming affect prompt-based learing? This project aims at analyzing this effect in stance classification. We train a stance classifier on the ibm stance classification dataset by fine-tuning an Alpaca model with a prompt and analyzing how does the selection of the few shots used in the prompt affect the performance of the model.

1. To evaluate the prompt-fine-tuning, run the following command

• Hyperparamter optimization

python scripts/run_prompt_fine_tuning.py --validate --optimize 

• Best Hyperparameters

python scripts/run_prompt_fine_tuning.py --validate --optimize 

2. To evaluate the in-context (prompt) setup run

python scripts/run_prompting.py --validate --optimize 

3. To evaluate DeBERTa (a normal classifier) with all hyperparameters, run the following

python scripts/optimize_baseline.py 

4. To evaluate Alpaca in a instructional tuning model run the following:

/run_prompt_fine_tuning.py --validate --optimize --alpaca

5. to evaluate Alpaca in using prompting you can run

/run_prompt_fine_tuning.py --validate --no-fine-tuning --alpaca

6. ro run the majority baseline

 python baseline.py --vast --majority --offline

The results of the experiments will be logged to your home directory. The parameters can be saved in config.yaml

Priming Sampling strategies

To run an experiment with a similar sampling strategy use the parameter --similar-examples. This will retrieve examples that are similar to each test instance. The similarity measure can be Sentence-Transformers --sentence-transformer. Example,

python scripts/run_prompting.py --validate --topic-similar --sentence-transformer

To run an experiment with a diversification sampling strategy use the parameter --diverse-examples. This will takes precomputed cluster centers of the training set as few shots. The number of clusters is then the few shot count provided in the configuration.

Topic Similarity

Examples on similar or diverse topics are sampled using a topic similarity, which relies on sentence transformer. the similarities can be used to apply the right sampling strategy.

1. To compute the similarities between all the validation and training arguments run the following

python scripts/run_develop_similarity_measure.py --compute-similarity --validate 

You have to specify the model used to compute the similarity to be --sentence-transformer

To load similarities from the code you can use the

similarities = load_similarities("ibmsc", "validation", similarity)

which returns a dictionary of dictionary where the indices of the first dictionary are the test indices and the indeices of the nested dictionary for each test index are the indices of the training set with the values being the similarity scores.

To find similar or diverse arguments for an argument in the validation or test set, you can use

examples = sample_diverse_examples(experiment, experiment_type, few_shot_size)

similar can be used for sample_similar

examples = sample_similar_examples(test_istance_index, similarities, df_training, few_shot_size)

Similar examples for both datasets are cached for efficiency and can be created by running the following commands python scripts/save_similar_examples.py --ks --topic-count --vast

Topic Diversification

To sample diverse examples, we use ward hierarchical clustering algorithm to cluster the trianing examples into $k$ cluster. The center of each cluster is then taken as an example. To find the diverse examples, we use the following command

python scripts/run_sample_diverse_examples --vast --validate --ibmsc 

This will precompute the cluster centers for k in [2, 4, 8, 16, 32, 64] To sample diverse examples, you can use the following command

examples = sample_diverse_examples(experiment, experiment_type, few_shot_size) where experiment is ibmsc or vast and experiment_type is validation or test.

Analysis

Mainly there are three types of analyses implemented on the prompting and instruction fine-tuning approaches.

Few shot size or training topic count effect on priming approaches

1. Run prompting or prompt-fine tuning approaches with anaylze k or analyze-topic-count as folows

python scripts/run_prompt_fine_tuning.py --analyze-k  --k 16 --seed 488

or

python scripts/run_prompt_fine_tuning.py --analyze-topic-count --topic-count 11 --seed 488

The ks or topic-counts are stored in conf.yaml and the analysis will be run for multiple seeds and the performance will be averaged and saved in the corresponding entry in /bigwork/nhwpajjy. To perform an analysis for one k or topic count instnance you can give it as a parameter using --k or --topic-count to produce the results for one seed you can specify the seed using --k

2. Store the path of the produced csv feel in conf.yaml analyze-k or analyze-topic-count

3. Draw the visualization using by running

python script/run_visualize_over_k_performance.py --k --topic-count --prompting --prompt-fine-tuning

Prime Anaylsis

To save the topically examples used for priming Alpaca, you can run the following. This will produce the test instances and the sampled claims in the following file. /bigwork/nhwpajjy/few-shot-priming/results/prime-analysis.tsv

python script/run_prompt_fine_tuning.py --analyze-primes